A dynamic separable convolution RCNN for lubrication condition identification of planetary roller screw mechanism

Abstract

Lubrication condition has a strong effect on the service ability of planetary roller screw mechanism (PRSM), so how to effectively identify the lubrication condition of PRSM is highly important in practical industrial applications. A dynamic separable convolution residual convolutional neural network (DSC-RCNN) method is proposed in this paper for the lubrication condition identification of PRSM. In the proposed method, a dynamic separable convolution (DSC) is developed by adopting depthwise separable convolution and dynamic convolution. To verify the learning competence of the proposed method, the PRSM failure test is carried out firstly and vibration data of the PRSM with and without grease are collected in multiple working conditions. Then, three experiments are implemented. The first one is to obtain the optimal number of the depthwise separable convolution in the DSC. The second one compares the effect of the DSC unit and the dynamic convolution unit on the diagnosis capacity of the proposed method. The last one compares SVM, BSA-SVM, AEs, LSSVM, LSTM, VGG-13, and the proposed method. The results reveal the best number of the depthwise separable convolution, the optimal unit of the proposed model and indicate that the DSC-RCNN has enormous recognition and transfer learning abilities.